Refresh
ActiveTraining gyms for computer use and software engineering work
About
Building RL environments for coding and computer use.
From their website
www.refresh.dev ↗Refresh builds simulation engines for coding and computer use to train AI coworkers that collaborate with humans. They partner with frontier labs and enterprises to create training environments and infrastructure for developing next-generation AI capabilities.
Refresh creates simulation engines that simulate coding and computer-use tasks, enabling labs and enterprises to train AI to operate like a human developer. The product involves building training gyms, datasets, and infrastructure to support AI learning and collaboration, including environments that teach AI to use the computer on terminals and GUI alike, with verifiable rewards integrated into simulations.
Who it’s for: Labs and enterprises developing advanced AI capabilities, particularly frontier laboratories and organizations building AI copilots or AI-assisted software development tools.
- simulation engines for coding and computer use
- training gyms for AI collaboration
- datasets and infrastructure for RL data operations
- environments to teach AI to operate terminals and GUI
- verifiable rewards integrated into simulations
- partnerships with frontier labs and enterprises
- tools to amplify human-AI collaboration
Open Roles listed, mentions of partnerships with frontier labs and enterprises, company focus on labs and training environments; indicates hiring and active development.
Founders · 2
Co-founder and CEO @ Refresh | ex-ML tech lead @ Uber AI, CS @ UIUC.
AmazonCo-founder and CTO @ Refresh | ex- @CapitalOne @Amazon / various startups (AI voice) a lot of data/scraping stuff. prev. computer vision research
Launch
High fidelity, realistic environments to train computer use and software engineering capabilities into LLMs
Refresh announces the creation of high-fidelity, web-based environments to train computer use and software engineering capabilities in frontier LLMs. They archive web content to generate task environments that reveal failure modes and develop mechanistic interpretability methods to elicit coding capabilities, aiming to assist teams training frontier intelligence.
Formerly “JupyPilot”, “Operative”, “Operative.sh”, “operative.sh”, “Operative”
Related startups
RL environments for long horizon AI Agents
Environments for Long Horizon Tool Use
